Electronic Theses and Dissertations (Masters)
Permanent URI for this collectionhttps://hdl.handle.net/10539/38006
Browse
3 results
Search Results
Item Generating Rich Image Descriptions from Localized Attention(University of the Witwatersrand, Johannesburg, 2023-08) Poulton, David; Klein, RichardThe field of image captioning is constantly growing with swathes of new methodologies, performance leaps, datasets, and challenges. One new challenge is the task of long-text image description. While the vast majority of research has focused on short captions for images with only short phrases or sentences, new research and the recently released Localized Narratives dataset have pushed this to rich, paragraph length descriptions. In this work we perform additional research to grow the sub-field of long-text image descriptions and determine the viability of our new methods. We experiment with a variety of progressively more complex LSTM and Transformer-based approaches, utilising human-generated localised attention traces and image data to generate suitable captions, and evaluate these methods on a suite of common language evaluation metrics. We find that LSTM-based approaches are not well suited to the task, and under-perform Transformer-based implementations on our metric suite while also proving substantially more demanding to train. On the other hand, we find that our Transformer-based methods are well capable of generating captions with rich focus over all regions of the image and in a grammatically sound manner, with our most complex model outperforming existing approaches on our metric suite.Item Analyzing the performance and generalisability of incorporating SimCLR into Proximal Policy Optimization in procedurally generated environments(University of the Witwatersrand, Johannesburg, 2024) Gilbert, Nikhil; Rosman, BenjaminMultiple approaches to state representation learning have been shown to improve the performance of reinforcement learning agents substantially. When used in reinforcement learning, a known challenge in state representation learning is enabling an agent to represent environment states with similar characteristics in a manner that would allow said agent to comprehend it as such. We propose a novel algorithm that combines contrastive learning with reinforcement learning so that agents learn to group states by common physical characteristics and action preferences during training. We subsequently generalise these learnings to previously encountered environment obstacles. To enable a reinforcement learning agent to use contrastive learning within its environment interaction loop, we propose a state representation learning model that employs contrastive learning to group states using observations coupled with the action the agent chose within its current state. Our approach uses a combination of two algorithms that we augment to demonstrate the effectiveness of combining contrastive learning with reinforcement learning. The state representation model for contrastive learning is a Simple Framework for Contrastive Learning of Visual Representations (SimCLR) by Chen et al. [2020], which we amend to include action values from the chosen reinforcement learning environment. The policy gradient algorithm (PPO) is our chosen reinforcement learning approach for policy learning, which we combine with SimCLR to form our novel algorithm, Action Contrastive Policy Optimization (ACPO). When combining these augmented algorithms for contrastive reinforcement learning, our results show significant improvement in training performance and generalisation to unseen environment obstacles of similar structure (physical layout of interactive objects) and mechanics (the rules of physics and transition probabilities).Item Learning to adapt: domain adaptation with cycle-consistent generative adversarial networks(University of the Witwatersrand, Johannesburg, 2023) Burke, Pierce William; Klein, RichardDomain adaptation is a critical part of modern-day machine learning as many practitioners do not have the means to collect and label all the data they require reliably. Instead, they often turn to large online datasets to meet their data needs. However, this can often lead to a mismatch between the online dataset and the data they will encounter in their own problem. This is known as domain shift and plagues many different avenues of machine learning. From differences in data sources, changes in the underlying processes generating the data, or new unseen environments the models have yet to encounter. All these issues can lead to performance degradation. From the success in using Cycle-consistent Generative Adversarial Networks(CycleGAN) to learn unpaired image-to-image mappings, we propose a new method to help alleviate the issues caused by domain shifts in images. The proposed model incorporates an adversarial loss to encourage realistic-looking images in the target domain, a cycle-consistency loss to learn an unpaired image-to-image mapping, and a semantic loss from a task network to improve the generator’s performance. The task network is con-currently trained with the generators on the generated images to improve downstream task performance on adapted images. By utilizing the power of CycleGAN, we can learn to classify images in the target domain without any target domain labels. In this research, we show that our model is successful on various unsupervised domain adaptation (UDA) datasets and can alleviate domain shifts for different adaptation tasks, like classification or semantic segmentation. In our experiments on standard classification, we were able to bring the models performance to near oracle level accuracy on a variety of different classification datasets. The semantic segmentation experiments showed that our model could improve the performance on the target domain, but there is still room for further improvements. We also further analyze where our model performs well and where improvements can be made.